Qwen3.6-27B-ThinkingCap-LoRA

This is a highly efficient, 141 MB Rank 64 GGUF LoRA adapter mathematically extracted from bottlecapai/ThinkingCap-Qwen3.6-27B.

Technical Details

• Method: SVD low-rank extraction of the delta between Qwen/Qwen3.6-27B and bottlecapai/ThinkingCap-Qwen3.6-27B.
• Target Modules: The extraction script dynamically isolated the specific deep layers (Layers 61, 62, 63) and the Multi-Token Prediction (MTP) head (mtp.layers.0) modified during the original finetune.
• Rank: 64
• Format: GGUF (qwen36-27b-thinkingcap-lora-rank64.gguf)

Usage (llama.cpp)

Load your preferred quant of the base model, and attach the adapter using the --lora-scaled flag (scaling at 1.0 is required as a baseline for SVD-extracted weights):

llama-server -m qwen3.6-27b-Q4_K_M.gguf --lora-scaled qwen36-27b-thinkingcap-lora-rank64.gguf 1.0